Merge pull request 'feat: GPU process list with per-process VRAM/utilization and kill' (#12) from feat/gpu-process-list into main

Reviewed-on: #12
2026-10-03 18:26:19 +00:00
commit e89135ef94
@@ -26,6 +26,7 @@ test: ## Run the test suite
 	$(UV) run python tests/test_security.py
 	$(UV) run python tests/test_rm_power.py
 	$(UV) run python tests/test_wireview.py
+	$(UV) run python tests/test_processes.py

 clean: ## Remove build artifacts
 	rm -rf frontend/dist frontend/node_modules
@@ -125,6 +125,8 @@ Notes:
 | `GET /api/gpu?gpu_index=0` | Name, driver version, VRAM |
 | `GET /api/dashboard?gpu_index=0` | Static dashboard info (VBIOS, CUDA cores, PCIe, BAR1, …). Live values come from the monitor WebSocket |
 | `GET /api/monitor?gpu_index=0` | One-shot monitoring snapshot: voltage, clocks, temp, power, fans, p-state, VRAM, utilization, throttle reasons |
+| `GET /api/processes?gpu_index=0` | Processes using this GPU, sorted by VRAM (descending): pid, user, dev, type (C/G/C+G), gpu_util_pct + mem_util_pct (last ~1 s only), vram_bytes, vram_pct, cpu_pct, mem_host_bytes, command |
+| `POST /api/processes/kill` | Body `{pid, signal: "TERM" \| "KILL", parent: false}` (default TERM) → send signal. `parent: true` signals the process's parent instead (response reports the signalled pid). Errors: `400` bad pid/signal or no parent, `404` process not found, `403` permission denied. **Note:** the pid need not be a GPU process — this is an arbitrary-pid signal primitive. The server runs as root, so with no users configured (open API) any reachable client can signal any process; enable auth before exposing the port on a shared system |

 ### Curve

@@ -9,6 +9,7 @@ import { CurveEditor } from "./components/CurveEditor/CurveEditor.js";
 import { PointTable } from "./components/PointTable/PointTable.js";
 import { PerformancePanel } from "./components/Limits/PerformancePanel.js";
 import { PerformanceMonitor } from "./components/Monitor/PerformanceMonitor.js";
+import { ProcessList } from "./components/Monitor/ProcessList.js";
 import { FanMonitor } from "./components/Monitor/FanMonitor.js";
 import { FanCurveEditor } from "./components/Fans/FanCurveEditor.js";
 import { WireViewPanel } from "./components/WireView/WireViewPanel.js";
@@ -290,16 +291,19 @@ function MainApp({
              )}
            </>
          ) : activeTab === "performance" ? (
-            <div className="flex gap-4 items-start w-full">
-              <div className="flex-1 min-w-0">
-                <PerformancePanel />
-              </div>
-              <div className="w-80 shrink-0 flex flex-col">
-                <PerformanceMonitor
-                  monitor={monitor}
-                  history={monitorHistory}
-                />
+            <div className="flex flex-col gap-4 w-full">
+              <div className="flex gap-4 items-start w-full">
+                <div className="flex-1 min-w-0">
+                  <PerformancePanel />
+                </div>
+                <div className="w-80 shrink-0 flex flex-col">
+                  <PerformanceMonitor
+                    monitor={monitor}
+                    history={monitorHistory}
+                  />
+                </div>
              </div>
+              <ProcessList />
            </div>
          ) : activeTab === "wireview" ? (
            <WireViewPanel
@@ -8,6 +8,7 @@ import type {
  FanState,
  FanPoint,
  DashboardInfo,
+  ProcessListResponse,
 } from "../types.js";

 export class ApiError extends Error {
@@ -119,6 +120,17 @@ export const api = {
  voltage: (gpuIndex: number) =>
    get<{ voltage_uv: number; voltage_mv: number }>("/voltage", gpuIndex),
  monitor: (gpuIndex: number) => get<MonitoringSample>("/monitor", gpuIndex),
+  processes: (gpuIndex: number) =>
+    get<ProcessListResponse>("/processes", gpuIndex),
+  killProcess: (
+    pid: number,
+    signal: "TERM" | "KILL" = "TERM",
+    parent = false,
+  ) =>
+    post<{ ok: boolean; pid: number; signal: string; parent: boolean }>(
+      "/processes/kill",
+      { pid, signal, parent },
+    ),
  snapshots: (gpuIndex: number) => get<SnapshotInfo[]>("/snapshots", gpuIndex),

  /** Write per-point frequency deltas. deltas: { pointIndex: deltaKhz } */
@@ -0,0 +1,254 @@
+import { useEffect, useRef, useState } from "react";
+import { Loader, X } from "lucide-react";
+import { api } from "../../api/client.js";
+import { useCurveStore } from "../../store/curveStore.js";
+import type { GpuProcess } from "../../types.js";
+import { fmt } from "../../utils/units.js";
+import { toast } from "sonner";
+import { ConfirmDialog } from "../common/ConfirmDialog.js";
+
+const POLL_MS = 3000;
+
+const TYPE_STYLES: Record<GpuProcess["type"], { label: string; cls: string }> =
+  {
+    C: { label: "C", cls: "bg-cyan-500/15 text-cyan-400 border-cyan-500/30" },
+    G: {
+      label: "G",
+      cls: "bg-emerald-500/15 text-emerald-400 border-emerald-500/30",
+    },
+    "C+G": {
+      label: "C+G",
+      cls: "bg-violet-500/15 text-violet-400 border-violet-500/30",
+    },
+  };
+
+const TH =
+  "px-3 py-1.5 font-semibold text-left text-[10px] uppercase tracking-wider text-zinc-500 whitespace-nowrap";
+const TD = "px-3 py-1.5 whitespace-nowrap";
+
+export function ProcessList() {
+  const { selectedGpuIndex } = useCurveStore();
+  const [processes, setProcesses] = useState<GpuProcess[] | null>(null);
+  const [memTotal, setMemTotal] = useState<number | null>(null);
+  const [error, setError] = useState<string | null>(null);
+  const [killTarget, setKillTarget] = useState<GpuProcess | null>(null);
+  const [pendingAction, setPendingAction] = useState<
+    "TERM" | "KILL" | "KILL_PARENT" | null
+  >(null);
+
+  // Ref so handleKill can trigger an immediate refresh after a kill.
+  const refreshRef = useRef<() => void>(() => {});
+
+  useEffect(() => {
+    let cancelled = false;
+
+    async function fetchOnce() {
+      try {
+        const data = await api.processes(selectedGpuIndex);
+        if (cancelled) return;
+        setProcesses(data.processes);
+        setMemTotal(data.mem_total_bytes);
+        setError(null);
+      } catch (e) {
+        if (!cancelled)
+          setError(e instanceof Error ? e.message : String(e));
+      }
+    }
+
+    refreshRef.current = () => {
+      void fetchOnce();
+    };
+    fetchOnce();
+    const timer = setInterval(fetchOnce, POLL_MS);
+    return () => {
+      cancelled = true;
+      clearInterval(timer);
+    };
+  }, [selectedGpuIndex]);
+
+  async function handleKill(action: "TERM" | "KILL" | "KILL_PARENT") {
+    if (!killTarget || pendingAction) return;
+    const pid = killTarget.pid;
+    setPendingAction(action);
+    try {
+      const signal = action === "TERM" ? "TERM" : "KILL";
+      const parent = action === "KILL_PARENT";
+      const res = await api.killProcess(pid, signal, parent);
+      toast.success(
+        parent
+          ? `Sent SIGKILL to parent process ${res.pid} of ${pid}`
+          : `Sent SIG${signal} to process ${pid}`,
+      );
+      setKillTarget(null);
+      refreshRef.current();
+    } catch (e) {
+      toast.error(e instanceof Error ? e.message : String(e));
+    } finally {
+      setPendingAction(null);
+    }
+  }
+
+  return (
+    <div className="bg-zinc-900 rounded-lg overflow-hidden flex flex-col">
+      {/* ── Header ─────────────────────────────────────────────────────── */}
+      <div className="flex items-center gap-2 px-3 py-2 border-b border-zinc-800">
+        <span className="text-xs text-zinc-500 uppercase tracking-wider font-semibold">
+          GPU Processes
+        </span>
+        {processes && processes.length > 0 && (
+          <span className="text-xs text-zinc-600">{processes.length}</span>
+        )}
+        <div className="ml-auto flex items-center gap-2">
+          {memTotal != null && (
+            <span className="text-xs text-zinc-600 font-mono">
+              VRAM total {fmt.bytes(memTotal)}
+            </span>
+          )}
+        </div>
+      </div>
+
+      {/* ── Error banner ────────────────────────────────────────────────── */}
+      {error && (
+        <div className="px-3 py-1.5 bg-red-900/40 border-b border-red-700 text-red-300 text-xs flex items-center justify-between">
+          <span>⚠ {error}</span>
+          <button
+            onClick={() => setError(null)}
+            className="ml-2 text-red-400 hover:text-red-200"
+          >
+            ✕
+          </button>
+        </div>
+      )}
+
+      {/* ── Table ───────────────────────────────────────────────────────── */}
+      {processes === null && !error ? (
+        <div className="p-8 flex items-center justify-center gap-2 text-zinc-500 text-xs">
+          <Loader size={14} className="animate-spin" />
+          Loading processes…
+        </div>
+      ) : processes && processes.length === 0 && !error ? (
+        <div className="p-8 text-center text-zinc-600 text-xs">
+          {memTotal === null
+            ? "GPU process data unavailable (NVML not initialized)."
+            : "No processes using this GPU."}
+        </div>
+      ) : (
+        <div className="overflow-x-auto">
+          <table className="w-full text-xs">
+            <thead>
+              <tr className="border-b border-zinc-800">
+                <th className={TH}>PID</th>
+                <th className={TH}>User</th>
+                <th className={TH}>Dev</th>
+                <th className={TH}>Type</th>
+                <th
+                  className={`${TH} text-right`}
+                  title="NVML per-process utilization — smoothed over the driver's sample window, so it lags instantaneous usage (accurate when steady)"
+                >
+                  GPU
+                </th>
+                <th className={`${TH} text-right`}>VRAM</th>
+                <th className={`${TH} text-right`}>VRAM %</th>
+                <th className={`${TH} text-right`}>CPU</th>
+                <th className={`${TH} text-right`}>MEM</th>
+                <th className={TH}>Command</th>
+                <th className={`${TH} text-right`}>Kill</th>
+              </tr>
+            </thead>
+            <tbody className="divide-y divide-zinc-800/60 font-mono">
+              {processes?.map((p) => (
+                <tr key={p.pid} className="hover:bg-zinc-800/40">
+                  <td className={TD}>{p.pid}</td>
+                  <td className={TD}>{p.user}</td>
+                  <td className={TD}>{p.dev}</td>
+                  <td className={TD}>
+                    <span
+                      className={`inline-block px-1.5 py-0.5 rounded border text-[10px] font-semibold ${TYPE_STYLES[p.type].cls}`}
+                      title={
+                        p.type === "C"
+                          ? "Compute"
+                          : p.type === "G"
+                            ? "Graphics"
+                            : "Compute + Graphics"
+                      }
+                    >
+                      {TYPE_STYLES[p.type].label}
+                    </span>
+                  </td>
+                  <td className={`${TD} text-right`}>
+                    {fmt.pct(p.gpu_util_pct ?? 0)}
+                  </td>
+                  <td className={`${TD} text-right`}>
+                    {fmt.bytes(p.vram_bytes)}
+                  </td>
+                  <td className={`${TD} text-right`}>
+                    {fmt.pct(p.vram_pct, 1)}
+                  </td>
+                  <td className={`${TD} text-right`}>
+                    {fmt.pct(p.cpu_pct, 1)}
+                  </td>
+                  <td className={`${TD} text-right`}>
+                    {fmt.bytes(p.mem_host_bytes)}
+                  </td>
+                  <td className="px-3 py-1.5 max-w-[320px] min-w-[160px]">
+                    <span
+                      className="block truncate text-zinc-300"
+                      title={p.command}
+                    >
+                      {p.command}
+                    </span>
+                  </td>
+                  <td className={`${TD} text-right`}>
+                    <button
+                      onClick={() => setKillTarget(p)}
+                      className="inline-flex items-center gap-1 px-2 py-1 rounded bg-zinc-800 hover:bg-red-900 text-zinc-400 hover:text-red-300 text-xs transition-colors"
+                      title={`Kill process ${p.pid}`}
+                    >
+                      <X size={11} />
+                      Kill
+                    </button>
+                  </td>
+                </tr>
+              ))}
+            </tbody>
+          </table>
+        </div>
+      )}
+
+      {/* ── Kill confirmation ───────────────────────────────────────────── */}
+      {killTarget && (
+        <ConfirmDialog
+          message={`Kill process ${killTarget.pid}?`}
+          detail={`${killTarget.command || "(unknown command)"} — owned by ${killTarget.user}, using ${fmt.bytes(
+            killTarget.vram_bytes,
+          )} VRAM. SIGTERM asks the process to exit; SIGKILL kills it immediately; SIGKILL Parent kills the parent process, which may take down the whole process family.`}
+          onCancel={() => setKillTarget(null)}
+          actions={[
+            {
+              label: pendingAction === "TERM" ? "Sending…" : "SIGTERM",
+              onClick: () => handleKill("TERM"),
+              disabled: pendingAction != null,
+              className:
+                "px-3 py-1.5 rounded text-xs font-semibold transition-colors bg-zinc-800 hover:bg-zinc-700 text-zinc-200",
+            },
+            {
+              label: pendingAction === "KILL" ? "Sending…" : "SIGKILL",
+              onClick: () => handleKill("KILL"),
+              disabled: pendingAction != null,
+              className:
+                "px-3 py-1.5 rounded text-xs font-semibold transition-colors bg-red-600 hover:bg-red-500 text-white",
+            },
+            {
+              label:
+                pendingAction === "KILL_PARENT" ? "Sending…" : "SIGKILL Parent",
+              onClick: () => handleKill("KILL_PARENT"),
+              disabled: pendingAction != null,
+              className:
+                "px-3 py-1.5 rounded text-xs font-semibold transition-colors bg-red-900 hover:bg-red-800 text-red-100",
+            },
+          ]}
+        />
+      )}
+    </div>
+  );
+}
@@ -1,10 +1,20 @@
+export interface DialogAction {
+  label: string;
+  onClick: () => void;
+  className?: string;
+  disabled?: boolean;
+}
+
 interface Props {
  message: string;
  detail?: string;
  confirmLabel?: string;
  isDestructive?: boolean;
-  onConfirm: () => void;
+  /** Required unless `actions` is provided (which replaces the single confirm button). */
+  onConfirm?: () => void;
  onCancel: () => void;
+  /** When provided, rendered after Cancel instead of the single confirm button. */
+  actions?: DialogAction[];
 }

 export function ConfirmDialog({
@@ -14,6 +24,7 @@ export function ConfirmDialog({
  isDestructive = false,
  onConfirm,
  onCancel,
+  actions,
 }: Props) {
  return (
    <div
@@ -33,17 +44,34 @@ export function ConfirmDialog({
          >
            Cancel
          </button>
-          <button
-            onClick={onConfirm}
-            className={[
-              "px-3 py-1.5 rounded text-xs font-semibold transition-colors",
-              isDestructive
-                ? "bg-red-600 hover:bg-red-500 text-white"
-                : "bg-emerald-600 hover:bg-emerald-500 text-white",
-            ].join(" ")}
-          >
-            {confirmLabel}
-          </button>
+          {actions ? (
+            actions.map((a) => (
+              <button
+                key={a.label}
+                onClick={a.onClick}
+                disabled={a.disabled}
+                className={[
+                  a.className ??
+                    "px-3 py-1.5 rounded bg-zinc-800 hover:bg-zinc-700 text-zinc-200 text-xs font-semibold transition-colors",
+                  a.disabled ? "opacity-40 cursor-not-allowed" : "",
+                ].join(" ")}
+              >
+                {a.label}
+              </button>
+            ))
+          ) : (
+            <button
+              onClick={() => onConfirm?.()}
+              className={[
+                "px-3 py-1.5 rounded text-xs font-semibold transition-colors",
+                isDestructive
+                  ? "bg-red-600 hover:bg-red-500 text-white"
+                  : "bg-emerald-600 hover:bg-emerald-500 text-white",
+              ].join(" ")}
+            >
+              {confirmLabel}
+            </button>
+          )}
        </div>
      </div>
    </div>
@@ -85,6 +85,26 @@ export interface GpuInfo {
  vram_gib: number | null;
 }

+export interface GpuProcess {
+  pid: number;
+  user: string;
+  dev: number;
+  type: "C" | "G" | "C+G";
+  // GPU utilization, only reported for processes active in the last ~1 s.
+  gpu_util_pct: number | null;
+  mem_util_pct: number | null;
+  vram_bytes: number | null;
+  vram_pct: number | null;
+  cpu_pct: number | null;
+  mem_host_bytes: number | null;
+  command: string;
+}
+
+export interface ProcessListResponse {
+  processes: GpuProcess[];
+  mem_total_bytes: number | null;
+}
+
 export interface SnapshotInfo {
  filepath: string;
  timestamp: string;
@@ -0,0 +1,237 @@
+"""GPU process list — which processes are using VRAM on a GPU.
+
+Per-PID VRAM and process type (compute/graphics) come from NVML; user,
+CPU usage, host memory and the full command come from /proc. The server
+runs as root, so it can read /proc entries of other users' processes.
+"""
+
+import os
+import pwd
+import time
+from typing import Any
+
+# NVML state lives in monitoring.py (nvmlInit is process-wide); import the
+# module so attribute lookups see the current values, not import-time copies.
+from . import monitoring as _mon
+
+_BOOT_TIME: float | None = None
+
+
+def _get_boot_time() -> float:
+    """System boot time as a Unix timestamp (cached; read from /proc/stat)."""
+    global _BOOT_TIME
+    if _BOOT_TIME is None:
+        _BOOT_TIME = time.time()  # fallback if /proc/stat is unreadable
+        try:
+            with open("/proc/stat") as f:
+                for line in f:
+                    if line.startswith("btime "):
+                        _BOOT_TIME = float(line.split()[1])
+                        break
+        except OSError:
+            pass
+    return _BOOT_TIME
+
+
+def _proc_info(pid: int) -> dict[str, Any] | None:
+    """Read user, CPU usage, host memory and command for a pid from /proc.
+
+    CPU% is the average over the process lifetime (total jiffies / age),
+    which needs a single /proc read — no sampling state. Returns None when
+    the process no longer exists.
+    """
+    base = f"/proc/{pid}"
+    try:
+        with open(f"{base}/stat", "rb") as f:
+            stat_raw = f.read().decode("ascii", "replace")
+        with open(f"{base}/statm") as f:
+            statm = f.read().split()
+        with open(f"{base}/cmdline", "rb") as f:
+            cmdline = f.read()
+    except OSError:
+        return None
+
+    # comm (field 2) may contain spaces and parentheses, so split at the
+    # last ')' — everything after it is fields 3..N.
+    lp = stat_raw.rfind(")")
+    if lp < 0:
+        return None
+    comm = stat_raw[stat_raw.index("(") + 1 : lp]
+    fields = stat_raw[lp + 2 :].split()
+    # field N -> fields[N-3]: utime=14, stime=15, starttime=22
+    if len(fields) < 20:
+        return None
+    try:
+        utime = int(fields[11])
+        stime = int(fields[12])
+        starttime = int(fields[19])
+    except ValueError:
+        return None
+
+    uid: int | None = None
+    try:
+        with open(f"{base}/status") as f:
+            for line in f:
+                if line.startswith("Uid:"):
+                    uid = int(line.split()[1])
+                    break
+    except OSError:
+        pass
+    try:
+        user = pwd.getpwuid(uid).pw_name if uid is not None else "unknown"
+    except KeyError:
+        user = str(uid)
+
+    hertz = os.sysconf("SC_CLK_TCK")
+    age_s = ((time.time() - _get_boot_time()) * hertz - starttime) / hertz
+    cpu_pct = (
+        round((utime + stime) / hertz / age_s * 100.0, 1) if age_s > 0 else None
+    )
+
+    page = os.sysconf("SC_PAGE_SIZE")
+    try:
+        mem_bytes = int(statm[1]) * page
+    except (IndexError, ValueError):
+        mem_bytes = None
+
+    parts = [p for p in cmdline.split(b"\0") if p]
+    command = " ".join(p.decode("utf-8", "replace") for p in parts) or comm
+
+    return {
+        "user": user,
+        "cpu_pct": cpu_pct,
+        "mem_bytes": mem_bytes,
+        "command": command,
+    }
+
+
+def _nvml_process_lists(handle: Any) -> dict[int, dict[str, Any]]:
+    """Map pid -> {"type": "C"|"G"|"C+G", "vram_bytes": int} for one GPU."""
+    out: dict[int, dict[str, Any]] = {}
+    pynvml = _mon._pynvml
+    try:
+        for p in pynvml.nvmlDeviceGetComputeRunningProcesses_v2(handle):
+            entry = out.setdefault(p.pid, {"type": "C", "vram_bytes": 0})
+            entry["vram_bytes"] = max(entry["vram_bytes"], p.usedGpuMemory or 0)
+    except pynvml.NVMLError:
+        pass
+    try:
+        for p in pynvml.nvmlDeviceGetGraphicsRunningProcesses_v2(handle):
+            entry = out.setdefault(p.pid, {"type": "G", "vram_bytes": 0})
+            entry["vram_bytes"] = max(entry["vram_bytes"], p.usedGpuMemory or 0)
+            if entry["type"] == "C":
+                entry["type"] = "C+G"
+    except pynvml.NVMLError:
+        pass
+    return out
+
+
+def _nvml_process_utilization(
+    handle: Any,
+) -> dict[int, tuple[float | None, float | None]]:
+    """Map pid -> (gpu_util_pct, mem_util_pct).
+
+    Only processes that used the GPU within the last ~1 s are reported
+    (driver >= 525; older drivers raise NOT_SUPPORTED -> empty map).
+    """
+    out: dict[int, tuple[float | None, float | None]] = {}
+    pynvml = _mon._pynvml
+    try:
+        for u in pynvml.nvmlDeviceGetProcessUtilization(handle, 0):
+            # Newer NVML renamed gpuUtil -> smUtil; accept both.
+            gpu_util = getattr(u, "smUtil", None)
+            if gpu_util is None:
+                gpu_util = getattr(u, "gpuUtil", None)
+            out[u.pid] = (
+                float(gpu_util) if gpu_util is not None else None,
+                float(u.memUtil) if u.memUtil is not None else None,
+            )
+    except pynvml.NVMLError:
+        pass
+    return out
+
+
+def list_gpu_processes(gpu_index: int = 0) -> dict[str, Any]:
+    """List processes using the given GPU, with per-PID details.
+
+    Returns {"processes": [...], "mem_total_bytes": int | None}. Processes
+    are sorted by VRAM usage (descending). Each process:
+    pid, user, dev (gpu_index), type ("C"/"G"/"C+G"), gpu_util_pct,
+    mem_util_pct, vram_bytes, vram_pct, cpu_pct, mem_host_bytes, command.
+    """
+    from .monitoring import get_vram_total
+
+    mem_total = get_vram_total(gpu_index)
+    processes: list[dict[str, Any]] = []
+
+    if _mon._NVML_AVAILABLE and _mon._nvml_initialized:
+        pynvml = _mon._pynvml
+        try:
+            handle = pynvml.nvmlDeviceGetHandleByIndex(gpu_index)
+            nvml_procs = _nvml_process_lists(handle)
+            util = _nvml_process_utilization(handle)
+        except pynvml.NVMLError:
+            nvml_procs, util = {}, {}
+
+        for pid, info in sorted(
+            nvml_procs.items(), key=lambda kv: kv[1]["vram_bytes"], reverse=True
+        ):
+            proc = _proc_info(pid)
+            if proc is None:
+                continue  # process exited between the NVML read and /proc read
+            gpu_util, mem_util = util.get(pid, (None, None))
+            vram = info["vram_bytes"]
+            processes.append(
+                {
+                    "pid": pid,
+                    "user": proc["user"],
+                    "dev": gpu_index,
+                    "type": info["type"],
+                    "gpu_util_pct": gpu_util,
+                    "mem_util_pct": mem_util,
+                    "vram_bytes": vram,
+                    "vram_pct": (
+                        round(vram / mem_total * 100.0, 1)
+                        if mem_total and vram is not None
+                        else None
+                    ),
+                    "cpu_pct": proc["cpu_pct"],
+                    "mem_host_bytes": proc["mem_bytes"],
+                    "command": proc["command"],
+                }
+            )
+
+    return {"processes": processes, "mem_total_bytes": mem_total}
+
+
+def kill_process(pid: int, sig: int) -> None:
+    """Send a signal to a pid.
+
+    Raises ProcessLookupError if the pid does not exist, PermissionError if
+    the caller may not signal it.
+    """
+    os.kill(pid, sig)
+
+
+def get_parent_pid(pid: int) -> int | None:
+    """Return the parent pid of a process.
+
+    Returns None when the process no longer exists, or 0 when it has no
+    parent (kernel threads, init).
+    """
+    try:
+        with open(f"/proc/{pid}/stat") as f:
+            raw = f.read()
+    except OSError:
+        return None
+    lp = raw.rfind(")")
+    if lp < 0:
+        return None
+    fields = raw[lp + 2 :].split()
+    # field 4 (ppid) -> fields[1]
+    if len(fields) < 2:
+        return None
+    try:
+        return int(fields[1])
+    except ValueError:
+        return None
@@ -9,6 +9,7 @@ Requires root (NvAPI needs it).
 import asyncio
 import logging
 import os
+import signal
 from contextlib import asynccontextmanager, suppress
 from pathlib import Path
 from typing import Any, Protocol
@@ -47,6 +48,7 @@ from .hal.monitoring import (
    shutdown_nvml,
    throttle_reasons_label,
 )
+from .hal.processes import get_parent_pid, kill_process, list_gpu_processes
 from .hal.ranges import get_clock_ranges
 from .hal.snapshot import (
    list_snapshots,
@@ -72,8 +74,8 @@ from .profiles.native import (
    rename_profile,
    save_profile,
 )
-from .wireview import create_device, find_wireview_ports
 from .safety import check_negative_freq_warnings, validate_write
+from .wireview import create_device, find_wireview_ports

 log = logging.getLogger("nvcurve.server")

@@ -837,6 +839,15 @@ class LoginRequest(BaseModel):
    password: str


+class ProcessKillRequest(BaseModel):
+    pid: int
+    # "TERM" (default) or "KILL".
+    signal: str = "TERM"
+    # When true, the signal is sent to the process's parent instead of the
+    # process itself (e.g. to stop a parent that keeps respawning the child).
+    parent: bool = False
+
+
 # ── Helper: run blocking HAL call in thread pool ──────────────────────────────


@@ -1107,6 +1118,77 @@ async def api_monitor(gpu_index: int = 0):
    return _sample_dict(sample)


+@app.get("/api/processes", tags=["GPU"], responses={**R_AUTH, **R_GPU})
+async def api_processes(gpu_index: int = 0):
+    """Processes using this GPU, sorted by VRAM (descending).
+
+    Per process: pid, user, dev (gpu_index), type (C/G/C+G), gpu_util_pct,
+    mem_util_pct, vram_bytes, vram_pct, cpu_pct, mem_host_bytes, command.
+    gpu_util_pct is only reported for processes active in the last ~1 s.
+    """
+    _require_gpu(gpu_index)
+    return await _run(list_gpu_processes, gpu_index)
+
+
+@app.post(
+    "/api/processes/kill",
+    tags=["GPU"],
+    responses={
+        **R_AUTH,
+        400: {
+            "description": "Invalid pid or signal (must be 'TERM' or 'KILL'), the process has no parent, or the target is init/kthreadd (pid 1/2)."
+        },
+        404: {"description": "Process (or its parent) not found."},
+        403: {"description": "Permission denied for the pid."},
+    },
+)
+async def api_process_kill(req: ProcessKillRequest):
+    """Send a signal to a process (or its parent).
+
+    Body: {pid, signal: 'TERM' | 'KILL', parent: bool}.
+    parent=true signals the process's parent instead of the process itself.
+    The response reports the pid that was actually signalled.
+    """
+    if req.pid <= 0:
+        raise HTTPException(status_code=400, detail="pid must be positive")
+    if req.signal not in ("TERM", "KILL"):
+        raise HTTPException(status_code=400, detail="signal must be 'TERM' or 'KILL'")
+    sig = signal.SIGKILL if req.signal == "KILL" else signal.SIGTERM
+
+    target = req.pid
+    if req.parent:
+        ppid = await _run(get_parent_pid, req.pid)
+        if ppid is None:
+            raise HTTPException(
+                status_code=404, detail=f"Process {req.pid} not found"
+            )
+        if ppid == 0:
+            raise HTTPException(
+                status_code=400, detail=f"Process {req.pid} has no parent"
+            )
+        target = ppid
+
+    # Never signal init (pid 1) or kthreadd (pid 2): SIGTERM to init is a
+    # system-shutdown vector, and kernel threads don't handle signals anyway.
+    if target in (1, 2):
+        raise HTTPException(
+            status_code=400,
+            detail=f"Refusing to signal pid {target} (init/kthreadd)",
+        )
+
+    try:
+        await _run(kill_process, target, sig)
+    except ProcessLookupError:
+        raise HTTPException(
+            status_code=404, detail=f"Process {target} not found"
+        ) from None
+    except PermissionError:
+        raise HTTPException(
+            status_code=403, detail=f"Permission denied for pid {target}"
+        ) from None
+    return {"ok": True, "pid": target, "signal": req.signal, "parent": req.parent}
+
+
@app.get("/api/snapshots", tags=["Snapshots"], responses=R_AUTH)
 async def api_snapshots():
    """List saved ClockBoostTable snapshots."""
@@ -2046,7 +2128,6 @@ async def api_shutdown():
    systems stop the service via systemd instead.
    """
    import os
-    import signal

    cfg: Config = _state["config"]
    if not cfg.allow_api_shutdown:
@@ -0,0 +1,346 @@
+"""Tests for the GPU process list + kill endpoints.
+
+Standalone (no pytest required):
+
+    python tests/test_processes.py
+
+Also works under pytest if available. Covers the /proc reader (including
+comm values with spaces and parentheses), the kill endpoint's input
+validation, a real SIGTERM round-trip, and the GET /api/processes payload.
+"""
+
+import contextlib
+import os
+import subprocess
+import sys
+import time
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
+
+from nvcurve import server  # noqa: E402
+from nvcurve.hal import processes  # noqa: E402
+
+PASS = 0
+FAIL = 0
+
+
+def check(name: str, cond: bool) -> None:
+    global PASS, FAIL
+    if cond:
+        PASS += 1
+        print(f"  PASS  {name}")
+    else:
+        FAIL += 1
+        print(f"  FAIL  {name}")
+
+
+def test_proc_info_self() -> None:
+    """_proc_info must return sane data for our own process."""
+    info = processes._proc_info(os.getpid())
+    check("self info not None", info is not None)
+    if info is None:
+        return
+    check(
+        "user is non-empty string",
+        bool(isinstance(info["user"], str) and info["user"]),
+    )
+    check(
+        "cpu_pct is None or non-negative number",
+        info["cpu_pct"] is None or (isinstance(info["cpu_pct"], float) and info["cpu_pct"] >= 0),
+    )
+    check("mem_bytes > 0", info["mem_bytes"] is not None and info["mem_bytes"] > 0)
+    check("command non-empty", isinstance(info["command"], str) and len(info["command"]) > 0)
+
+
+def test_proc_info_missing() -> None:
+    """A pid that does not exist must yield None (not an exception)."""
+    check("missing pid -> None", processes._proc_info(99999999) is None)
+
+
+def test_proc_info_comm_with_parens() -> None:
+    """comm containing spaces and parentheses must not break /proc/stat parsing.
+
+    The kernel comm is set via prctl(PR_SET_NAME); `exec -a` only changes
+    argv[0] and would leave comm as the executable name, so the parens path
+    would never be exercised.
+    """
+    proc = subprocess.Popen(
+        [
+            sys.executable,
+            "-c",
+            "import ctypes, time\n"
+            "ctypes.CDLL(None).prctl(15, b'proc (test) x', 0, 0, 0)\n"
+            "time.sleep(30)",
+        ]
+    )
+    try:
+        # Wait until /proc/<pid>/stat exists and comm actually carries the
+        # parenthesised name (prctl may take a moment after exec).
+        comm = ""
+        for _ in range(100):
+            try:
+                with open(f"/proc/{proc.pid}/stat") as f:
+                    raw = f.read()
+                comm = raw[raw.index("(") + 1 : raw.rfind(")")]
+            except OSError:
+                pass
+            if "(" in comm:
+                break
+            time.sleep(0.05)
+        check("fixture comm contains parens", "(" in comm)
+        info = processes._proc_info(proc.pid)
+        check("weird-comm info not None", info is not None)
+        if info is not None:
+            check("weird-comm command parsed", "prctl" in info["command"])
+    finally:
+        proc.kill()
+        proc.wait()
+
+
+def test_kill_endpoint_validation() -> None:
+    """The kill endpoint must reject bad input and unknown pids."""
+    from fastapi.testclient import TestClient
+
+    client = TestClient(server.app)
+    r = client.post(
+        "/api/processes/kill", json={"pid": os.getpid(), "signal": "BOGUS"}
+    )
+    check("invalid signal -> 400", r.status_code == 400)
+    r = client.post("/api/processes/kill", json={"pid": -5, "signal": "TERM"})
+    check("negative pid -> 400", r.status_code == 400)
+    r = client.post("/api/processes/kill", json={"pid": 0, "signal": "TERM"})
+    check("zero pid -> 400", r.status_code == 400)
+    r = client.post("/api/processes/kill", json={"pid": 99999999, "signal": "TERM"})
+    check("unknown pid -> 404", r.status_code == 404)
+
+
+def test_kill_endpoint_real() -> None:
+    """A real SIGTERM round-trip: the process must actually exit."""
+    from fastapi.testclient import TestClient
+
+    client = TestClient(server.app)
+    proc = subprocess.Popen(["sleep", "30"])
+    try:
+        r = client.post(
+            "/api/processes/kill", json={"pid": proc.pid, "signal": "TERM"}
+        )
+        check("kill -> 200", r.status_code == 200)
+        check("response ok", r.json().get("ok") is True)
+        try:
+            proc.wait(timeout=5)
+            check("process exited", True)
+        except subprocess.TimeoutExpired:
+            check("process exited", False)
+    finally:
+        proc.kill()
+        proc.wait()
+
+
+def test_kill_endpoint_refuses_init_kthreadd() -> None:
+    """The endpoint must refuse to signal init (pid 1) or kthreadd (pid 2).
+
+    SIGTERM to init is a system-shutdown vector; kernel threads don't handle
+    signals anyway. This fires before any NVML/state access, so no GPU needed.
+    """
+    from fastapi.testclient import TestClient
+
+    client = TestClient(server.app)
+    for pid in (1, 2):
+        r = client.post(
+            "/api/processes/kill", json={"pid": pid, "signal": "KILL"}
+        )
+        check(f"pid {pid} refused -> 400", r.status_code == 400)
+
+
+def test_get_parent_pid() -> None:
+    """get_parent_pid must return the kernel ppid and degrade gracefully."""
+    check("self ppid matches", processes.get_parent_pid(os.getpid()) == os.getppid())
+    check("missing pid -> None", processes.get_parent_pid(99999999) is None)
+    check("pid 1 has no parent", processes.get_parent_pid(1) == 0)
+
+
+def test_kill_parent_endpoint() -> None:
+    """parent=true must signal the parent, leaving the child alive."""
+    from fastapi.testclient import TestClient
+
+    client = TestClient(server.app)
+
+    # Validation cases first (no fixture needed).
+    r = client.post(
+        "/api/processes/kill", json={"pid": 99999999, "signal": "KILL", "parent": True}
+    )
+    check("parent of missing pid -> 404", r.status_code == 404)
+    r = client.post(
+        "/api/processes/kill", json={"pid": 1, "signal": "KILL", "parent": True}
+    )
+    check("pid 1 has no parent -> 400", r.status_code == 400)
+
+    # bash forks a background sleep (child), then sleeps itself (parent).
+    parent = subprocess.Popen(["bash", "-c", "sleep 30 & sleep 30"])
+    child: int | None = None
+    for _ in range(100):
+        try:
+            with open(f"/proc/{parent.pid}/task/{parent.pid}/children") as f:
+                kids = f.read().split()
+            if kids:
+                child = int(kids[0])
+                break
+        except OSError:
+            pass
+        time.sleep(0.05)
+    try:
+        check("fixture has a child", child is not None)
+        if child is None:
+            return
+        r = client.post(
+            "/api/processes/kill",
+            json={"pid": child, "signal": "KILL", "parent": True},
+        )
+        check("kill parent -> 200", r.status_code == 200)
+        check("response targets the parent", r.json().get("pid") == parent.pid)
+        check("response reports parent flag", r.json().get("parent") is True)
+        try:
+            parent.wait(timeout=5)
+            check("parent died", True)
+        except subprocess.TimeoutExpired:
+            check("parent died", False)
+        time.sleep(0.2)
+        try:
+            os.kill(child, 0)
+            check("child survived (orphaned)", True)
+        except ProcessLookupError:
+            check("child survived (orphaned)", False)
+    finally:
+        for victim in (parent,):
+            with contextlib.suppress(ProcessLookupError, subprocess.TimeoutExpired):
+                victim.kill()
+                victim.wait(timeout=5)
+        if child is not None:
+            with contextlib.suppress(ProcessLookupError):
+                os.kill(child, 9)
+
+
+def test_processes_endpoint() -> None:
+    """GET /api/processes must return the HAL payload for a known GPU."""
+    from fastapi.testclient import TestClient
+
+    fake = {
+        "processes": [
+            {
+                "pid": 1,
+                "user": "root",
+                "dev": 3,
+                "type": "C",
+                "gpu_util_pct": 42.0,
+                "mem_util_pct": 1.0,
+                "vram_bytes": 1000,
+                "vram_pct": 0.1,
+                "cpu_pct": 1.5,
+                "mem_host_bytes": 2048,
+                "command": "/usr/bin/fake",
+            }
+        ],
+        "mem_total_bytes": 1000000,
+    }
+    orig = server.list_gpu_processes
+    server._state["gpus"][3] = {"gpu": object(), "gpu_name": "fake"}
+    server.list_gpu_processes = lambda idx: fake
+    try:
+        client = TestClient(server.app)
+        r = client.get("/api/processes", params={"gpu_index": 3})
+        check("processes -> 200", r.status_code == 200)
+        check("payload matches HAL", r.json() == fake)
+        r = client.get("/api/processes", params={"gpu_index": 99})
+        check("unknown gpu -> 404", r.status_code == 404)
+    finally:
+        server.list_gpu_processes = orig
+        server._state["gpus"].pop(3, None)
+
+
+def test_list_gpu_processes_shape() -> None:
+    """With NVML available, list_gpu_processes must return a well-shaped payload."""
+    from nvcurve.hal import monitoring
+
+    if not monitoring._NVML_AVAILABLE:
+        print("  SKIP  NVML not available")
+        return
+    if not monitoring.init_nvml():
+        print("  SKIP  NVML init failed (no driver?)")
+        return
+    try:
+        data = processes.list_gpu_processes(0)
+        check("has processes list", isinstance(data.get("processes"), list))
+        check("has mem_total_bytes", "mem_total_bytes" in data)
+        keys = {
+            "pid",
+            "user",
+            "dev",
+            "type",
+            "gpu_util_pct",
+            "mem_util_pct",
+            "vram_bytes",
+            "vram_pct",
+            "cpu_pct",
+            "mem_host_bytes",
+            "command",
+        }
+        for p in data["processes"]:
+            if not keys.issubset(p.keys()):
+                check(f"process {p.get('pid')} has all keys", False)
+                break
+        else:
+            check("all processes have all keys", True)
+        # Sorted by VRAM descending.
+        vram = [p["vram_bytes"] or 0 for p in data["processes"]]
+        check("sorted by VRAM desc", vram == sorted(vram, reverse=True))
+    finally:
+        monitoring.shutdown_nvml()
+
+
+def test_list_gpu_processes_no_nvml() -> None:
+    """With NVML unavailable, list_gpu_processes returns an empty payload.
+
+    This is the branch the shape test skips (no driver): the UI uses
+    mem_total_bytes is None to render 'data unavailable' instead of
+    'no processes'.
+    """
+    from nvcurve.hal import monitoring
+
+    orig_avail = monitoring._NVML_AVAILABLE
+    orig_init = monitoring._nvml_initialized
+    monitoring._NVML_AVAILABLE = False
+    monitoring._nvml_initialized = False
+    try:
+        data = processes.list_gpu_processes(0)
+        check("empty processes list", data["processes"] == [])
+        check("mem_total_bytes is None", data["mem_total_bytes"] is None)
+    finally:
+        monitoring._NVML_AVAILABLE = orig_avail
+        monitoring._nvml_initialized = orig_init
+
+
+def main() -> None:
+    tests = [
+        test_proc_info_self,
+        test_proc_info_missing,
+        test_proc_info_comm_with_parens,
+        test_kill_endpoint_validation,
+        test_kill_endpoint_real,
+        test_kill_endpoint_refuses_init_kthreadd,
+        test_get_parent_pid,
+        test_kill_parent_endpoint,
+        test_processes_endpoint,
+        test_list_gpu_processes_shape,
+        test_list_gpu_processes_no_nvml,
+    ]
+    for t in tests:
+        print(f"== {t.__name__} ==")
+        t()
+    print()
+    print(f"{PASS} passed, {FAIL} failed")
+    if FAIL:
+        sys.exit(1)
+
+
+if __name__ == "__main__":
+    main()